Chongfan Technology

News

All Categories >

Home page

AI Visual Algorithm Model Customization Service

2026

AI Visual Algorithm Model Customization Service

Author:

I. Project Background

Customer : A leading automotive parts supplier
Scene : Full-inspection station for automotive bearing appearance
Demand : Replacing manual visual inspection, it enables automated detection of 23 types of defects on bearing outer rings, end faces, and inner walls, while supporting mixed‑model production across more than 20 bearing types.

AI Vision Inspection in Manufacturing: Real-Time Defect Detection & Quality Control

II. Core Pain Points

Table

Pain point	Current status of manual operations	Customization Requirements
The defect is minor.	Edge curling, slight rust—easy to overlook with the naked eye due to visual fatigue.	Pixel-level segmentation, with a detection rate of ≥99.5%
Many models available	More than 20 models have significant size variations, requiring retraining when switching between them.	A rapid adaptation of an algorithm library, with no need to switch programs.
Speed bottleneck	Manual full inspection: 600 pcs/h, unable to keep up with the production line’s takt time.	Detection speed ≥1200 pcs/h
Data scarcity	With a small number of defective samples, traditional deep learning requires tens of thousands of labeled images.	Small Sample + Transfer Learning: Cold Start with Hundreds of Samples

III. Customized Solution (Four-Phase Delivery)

Phase 1: Customization of Optics and Data Acquisition

To address the interference caused by metallic reflections from bearings and the presence of anti-rust oil on their surfaces, a design was developed. Standing rotational photography + dual-axis drive The solution eliminates flat‑field pixel distortion and secondary contamination, while simultaneously acquiring both normal and defective samples to build a structured dataset.

Phase 2: Customized Training of Algorithmic Models

Based on a self-developed training framework, it integrates OCR + Object Detection + Semantic Segmentation Multitask model:

Split branch : High-resolution preservation, precise extraction of crack and crush‑induced edge boundaries;
Detection branch : YOLO cascade, for rapid detection of dirt and dents;
Transfer learning : A general-purpose pre-trained model for metal surfaces; the new version requires only 50–100 additional samples for fine-tuning.

AI-Based Industrial Defect Detection Using Deep Learning

Baidu AI Studio – A One-Stop AI Development and Training Platform

Phase 3: Edge Deployment and System Integration

After model quantization, the model is deployed on industrial PCs or edge computing boxes on the production line, achieving inference for a single frame in under 0.2 seconds. OK/NG results and defect coordinates are reported back via the MES interface, enabling data traceability.

Phase 4: Continuous Optimization and Iteration

After going live, the model undergoes incremental learning based on feedback from false positives and missed detections, enabling version updates without any downtime.

IV. Key Technical Indicators

Table

Indicator	Delivery Result
Detection speed	1200 pcs/h (single station)
Defect Detection Rate	≥ 99.59%
Inspection pass rate	≤ 5%
Single-frame processing time	< 0.2 s
Changeover adaptation time	< 30 min (same material)
Labor substitution	The number of full-inspection positions has been reduced by approximately 80%.